ADVANCES in NATURAL and APPLIED SCIENCES

Transcription

1 ADVANCES in NATURAL and APPLIED SCIENCES ISSN: Published BYAENSI Publication EISSN: May 11(7): pages Open Access Journal Feature Extraction And Classification Algorithm For Screening Of Melanoma 1 O. Vishnuvarthini and 2 Dr. B. Sathiyabhama 1 PG Scholar-Computer Science and Engineering Sona College of Technology Salem, India. 2 Professor/Head Computer Science and Engineering Sona College of Technology Salem, India. Received 28 January 2017; Accepted 22 May 2017; Available online 28 May 2017 Address For Correspondence: O.Vishnuvarthini, PG Scholar-Computer Science and Engineering Sona College of Technology Salem, India. vishnuvarthini02@gmail.com Copyright 2017 by authors and American-Eurasian Network for ScientificInformation (AENSI Publication). This work is licensed under the Creative Commons Attribution International License (CC BY). ABSTRACT The proper identification of melanomas in dermoscopy images is still an up to date challenge. This paper addresses two different systems for the detection of melanoma in dermoscopy images. The first system uses global methods to classify skin lesions, whereas the second system uses local features and the bag-of-features classifier. This present work proposes to determine the best system for skin lesion classification. Global features are using color descriptors and texture features to perform the classification. The texture features are extracted using GLCM matrix and the color features are extracted by computing histogram in five color spaces. Local features are using SURF descriptors and BOF classifier to perform the classification. It is concluded that color features outperform texture features when used alone and that both methods achievevery good results, i.e., Sensitivity is 86 percent and Specificity is 100 percent for global methods against Sensitivity is 90percentand Specificity is 100percentfor local methods. KEYWORDS: Melanoma, Color, Texture, Feature extraction INTRODUCTION Melanoma is one of the deadliest forms of cancer. It can occur in any part of the body that contains melanocytes. The main cause of melanoma is excessive exposure to ultraviolet (UV) radiation reaching the skin. UV rays from the sun and other sources (such as tanning booths) can damage skin cells, causing the cells to grow abnormally. Fair-skinned people have a higher risk for melanoma and other types of skin cancer if they have too much exposure to sunlight. It is necessary to control and diagnose at very early stage to save from death globally. A technique used by dermatologist to diagnose skin lesions is dermoscopy. Dermoscopy is a non-invasive examination technique based on the use of incident light and oil immersion to make possible the visual examination of sub surface structures of the skin.since differential diagnosis of melanoma from melanocytic nevi is often not straight forward especially in the early stage, the diagnostic accuracy of dermoscopy is also depending on the training of the dermatologist. So that automatic diagnosis is essential tool for less experienced physicians. It considered to be a double reading system were physicians take into consideration the information provided by computer before making decision. Computer is not more intelligent than human but it may be able to extract some information, such as texture features, that may not be readily perceived by human eyes. Several scoring systems and algorithms such as the ABCD rule for epiluminescence, the seven-point checklist, and the Menzies method[4], [5]have been proposed to improve the diagnostic performance of less experienced clinicians. Most of the proposed techniques require segmentation process that considers being a fatal problem due to the irregularity of the tumor, where dermoscopy views of histological tissues show structures mostly arranged in a variety of patterns. So that, ToCite ThisArticle: O. Vishnuvarthini and. B. Sathiyabhama., Feature Extraction And Classification Algorithm For Screening Of Melanoma. Advances in Natural and Applied Sciences. 11(7); Pages:

2 615 O. Vishnuvarthini and Dr. B. Sathiyabhama., 2017/Advances in Natural and Applied Sciences. 11(7) May 2017, Pages: automatic segmentation of different structures, like nuclei, cytoplasm, vessels etc., is difficult, and cannot be done in general approach[6]-[10]. Classification of dermoscopy images is an image processing problem. Several approaches have been proposed to deal with various data sets. The solutions are classified into two different categories. Using global features and local features. Global features (color pigments and gradient histogram, size, shape, etc..) are extracted from the given images and a model is built on top of that used for the classification of melanoma images. However these setting fail to recognize objects in more complex settings e.g., when the object shape, color and texture exhibit severe changes and cannot be easilydescribed by global measurements extracted from the image. But the recent studies are discussion about the local features of images, small regions of images and describing each region by a set of local features. This method does not require global model for the object and can easily cope with changes in shape, color and texture, achieving good results in very challenging conditions. Related Work: M. sheha, M. S.Mabrouk and A. Sharawy propose a features extacted are based on gray level Co-occurence matrix (GLCM) and using Multilayer perceptron classifier (MLP) to classify between melanocytic nevi and malignant melanoma. ANN is considered as important way for classification and the ANN classification procedure used is MLP. MLP classifier was proposed with two different techniques in training and testing process: Automatic MLP and Traditional MLP. Results indicated that texture analysis method is a useful method for discrimination of melanocytic skin tumours with high accuracy. Segmentation process is avoided. The method uses the steps of Pre-processing, feature extraction, feature selection, classification and then evaluation. The results indicated that Traditional MLP yielded the better performance when compared to the Automatic MLP. F.S. Khan, J. Van de Weijer and M. Vanrell proposes a new approach[2] to analyse the problem of object recognition within the bag-of-words framework using multiple cues, particularly combining shape and color information. In order to combine multiple cues within the bag-of-words framework, we consider two properties that are especially desirable for final image representation: feature binding and vocabulary compactness. Feature binding involves combining information from different features at the local level and not at image level. Vocabulary compactness denotes to the property of having a separate vocabulary for each of the different cues. Do not handle the issue of multiple cues. There exist two main approaches are, Early fusion fusing local descriptors together and late fusion concatenates histogram representation of both color and shape. M. Silveria, J. Nascimento, J. Marques, A. Marcal, T. Mendonca, S. Yamauchi and J. Maeda,presents [3] a methods such as adaptive thresholding (AT), adaptive snake (AS), EM level set (EM-LS) and fuzzy based splitand-merge algorithm (FBSM) for the segmentation of skin lesion in dermoscopic images with an increase of gross errors. The best results were obtained by the AS and EM-LS methods, which are semi-supervised methods. The best fully automatic method was FBSM, with results only slightly worse than AS and EM-LS which are more robust than other methods since the number of gross segmentation errors is smaller. Recognition with local features [5] provides local features can give efficient representations suitable for robust object recognition. SVM and kernel methods have begun to be used for appearance based object recognition. SVM gave significant increases in performance, which again confirms the advantage of largemargin classifiers regardless of the underlying data representation. Recognition is based on the computation of the similarity between two invariant vectors. Matching is performed on discriminant points of an image and a standard voting algorithm is used to find the closest model to an image. Automated diagnosis of PSL proposes digital dermoscopy [6] analyser which is used to evaluate a series of clinically typical, flat pigmented skin lesions. Difficult to differentiate early melanoma from other PSLs. ANN provides robust performance. It combines the reliability of automatic digital image processing of the DB Dermo-Mips system makes the classification method adaptable. Proposed System: A. Segmentation: We have developed texture level segmentation using entropy filtering technique. And to minimize the time required for image analysis. Image segmentation is the process of partitioning a digital image into multiple segments. The goal of segmentation is to simplify or change the representation of an image into something that is more meaningful and easier to analyse. Segmentation divides the image into its constituent regions or objects. Texture based segmentation follows this basic procedure 1) Compute a segmentation function. This is an image whose dark regions are the objects you are trying to segment. 2) Compute foreground markers. These are connected blobs of pixels within each of the objects. 3) Compute background markers. These are pixels that are not part of any object. 4) Based on that analysis can be evaluated along with threshold values. Then the images are segmented.

3 616 O. Vishnuvarthini and Dr. B. Sathiyabhama., 2017/Advances in Natural and Applied Sciences. 11(7) May 2017, Pages: B. Global Features extraction: In this paper we are discussing about both the approaches and comparing the results. The first system describes the dermoscopy image by a set of global features and uses a classifier to classify melanomas from non-melanoma images. In this approach we are used five different color spaces (RGB, LUV, OPP, HSV) and texture features for the classification. There are 8 texture features are calculated using the GLCM matrix. We tried to build 6 individual classifiers to build using 5 different color spaces and texture features. In the other experiment we built two different classifier by combining texture features + RGB and Texture features + HSV features to understand how the accuracy is affecting by combining texture and color. The results are discussed in the following sections. This is a supervised system since the classifier learns to detect the melanoma lesions using a training set of images which is labelled by an expert. Each image I is characterized by a feature vector Vi, and assign to a binary label Vi = {0, 1}. The classifier trained to classify both types of images using the training data. C. Color Features: The most popular features used in demoscopy analysis are color statistics such as the mean color and color variance. The doctors are using color of the lesion to classify melanoma or non-melanoma. Most of the studies are talking about RGB color space. The RGB color space represents a target color as a mixture of three primary colors, Red, Green, and Blue. However this color space has a number of drawbacks, it is not perceptually uniform, it depends on the acquisition setup, and it exhibits a high correlation among the three color channels. To overcome these difficulties, other color representations have been proposed. For this work we are using five different color spaces, RGB, HSV, OPP, LUV, LAB. D. Texture Features: Texture is used in identifying objects or regions of interest in an image. Texture contains important information about the structural arrangement of surfaces. Haralick introduced the use of co-occurrence probabilities using GLCM for extracting various texture features which contains information about image texture characteristics. Some of the texture features extracted and they are Energy, Entropy, Contrast, Homogeneity, Correlation, Dissimilarity, Inverse Difference Moment and Maximum Probability. E. Local Features Extraction: The second approach using BOF approach. First a set of key points is selected insides the lesion region. Then each key point characterized by a vector of local features. This feature vector represents color and texture properties in a local patch cantered at the key point. Since the number of key points and local features varies from image to image, we cannot directly feed a classifier with these data. Instead all local features associated with all the training images are gathered and used to compute a smaller set of prototypes (centroids) denoted as visual words. Then the local features of each dermoscopy images are assigned to the nearest visual words, and a histogram is computed. The histogram counts the no of times each visual word was selected A statistical classifier is then trained to discriminate melanoma lesions from the non-melanoma ones, using the histogram of visual words as input. F. SURF: Speed Up Robust Features (SURF) is a local feature detector and descriptor that can be used for tasks such as object recognition. It is partly inspired by the SIFT descriptor. SURF is faster and good at handling image with blurring or rotation. SURF algorithm has three main parts: interest point detection, local neighbourhood description and matching. It uses a blob detector based on the Haessian matrix as follows: Where Lxx(X,Y,σ) is the laplacian of Gaussian of the image and which are second order derivatives of the gray scale image. Classification: The classification system consists of three processing steps. First the skin lesion is segmented. Then texture and color features are extracted. Finally a SVM classifier is used to perform the classification. A. Classification by Global Features: In Color descriptor, the given image is converted into feature vector using different color spaces. The size of the feature vector is 256 for each color space. There are 5 different classifiers are built to understand the classification accuracy. We have tried with multiple kernels in SVM but linear kernel outperforms other kernels. The following table shows the accuracy of the classifiers.

4 617 O. Vishnuvarthini and Dr. B. Sathiyabhama., 2017/Advances in Natural and Applied Sciences. 11(7) May 2017, Pages: Table 1: Color Space Results RGB HSV LUV OPP LAB In the second experiment we have used texture descriptor to perform the classification. In texture descriptor the given image is converted into a GLCM matrix and some of the texture features are extracted and will form the feature vector. The size of the feature vector is 8 and RBF kernel is given more accuracy than the linear kernel. Table 2: Texture Features Results Texture In the third experiment we have combining the texture descriptor with color descriptor to understand the performance of the classification. We have built two different classifiers for texture + RGB and texture + HSV color space. Table 3: Texture + Color Space Results Texture RGB Texture+ HSV In combining RGB with texture features and size of feature vector is 776 and linear kernel is outperformed the RBF kernel and obtained the confusion with sensitivity and specificity SE = 65% and SP = 80%. B. Classification by Local Features: In local features we have created 500 visual vocabularies to perform the SVM classification. By looking the following results it evident the local features are performing good to identify melanoma images from nonmelanoma images with better accuracy. Table 4: Local Features Result SURF Performance Evaluation Sensitivity Accuracy Specificity Fig. 1: Performance Evaluation By seeing the above figure (Fig, 1), its evident that local feature based classification is outperform others. RGB., HSV are producing the same accuracy and comparable with the local features. Texture alone is not

5 618 O. Vishnuvarthini and Dr. B. Sathiyabhama., 2017/Advances in Natural and Applied Sciences. 11(7) May 2017, Pages: performing well but when we are combining texture with color descriptors we are getting slightly better accuracy. In local feature descriptor we have generated 500 visual words. We have done some experiments by increasing and decreasing the no of visual words and we have seen accuracy difference. Conclusion: In this paper we compared two different methods for the detection of melanomas in dermoscopy images based on local and global features. It was concluded that color features perform better than texture features alone achieving classification scores SE=86.2% and SP=100%. Concerning the use of global features versus local features both systems achieving good results. However, the local system performs better in terms of classification by producing SE=89.7 and SP=100%. REFERENCES 1. sheha, M., M.S. Mabrouk and A. Sharawy, Automatic detection of melanoma skin cancer using texture analysis, Int. J. Comput. Appl., 42(20): Khan, F.S., J. Van de Weijer and M. Vanrell, Top down color attention for object recognition in Proc. IEEE 12 th Int. Conf. Comput. Vis., pp: Silveria, M., J. Nascimento, J. Marques, A. Marcal, T. Mendonca, S. Yamauchi and J. Maeda, Comparison of segmentation methods for melanoma diagnosis in dermoscopy images, IEEE J. Sel. Topics Signal Processing., 3(1): Iyatomi, H., H. Oka, M.E. Celebi, M. Hashimoto, M. Hagiwara, M. Tanaka and K. Ogawa, An Internet-based Melanoma Diagnostic System Toward the Practical Application. 5. Walravem, C., B. Caputo and A. Graf, Recognition with local features: The Kernel recipe, in Proc. 9 th IEEE Int. Conf. Comput. Vis., pp: Rubegni, P., G. Cevenini, M. Burroni, R. Perotti, G. Dell Eva, P. Sbano and C. Miracco, Automated diagnosis of pigmented skin lesions, Int. J. Cancer, 101(6): Abbas, Q., M.E. Celebi, C. Serrano, I.F. García, G. Ma, Pattern classification of dermoscopy images: a perceptually uniform model. Pattern Recogn., 46: Argenziano, G., G. Fabbrocini, P. Carli, V. De Giorgi, E. Sammarco, M. Delfino, Epiluminescence microscopy for the diagnosis of doubtful melanocytic skin lesions. Comparison of the ABCD rule of dermatoscopy and a new 7-point checklist based on pattern analysis. Arch. Dermatol., 134: A Bag-of-Features Approach for the Classification Argenziano, G., H. Soyer, V. De Giorgi, P. Carli, M. Delfino, A. Ferrari, R. Hofmann-Wellenhof, D. Massi, G. Mazocchetti, M. Scalvenzi, I. Wolf, Interactive atlas of dermoscopy. Edra Medical Publishing and New Media, Milan Arivazhagana, S., L. Ganesanb, S.P. Priyala, 2006.Texture classification using gabor wavelets based rotation invariant features. Pattern Recogn. Lett., 27: Barata, C., J.S. Marques, J. Rozeira, A system for the detection of pigment network in dermoscopy images using directional filters. IEEE Trans. Biomed. Eng. 59(10): Bratkova, M., S. Boulos, P. Shirley, orgb: A pratical opponent color space for computer graphics. IEEE Comput. Graphics Appl., 29: Burges, C.J.C., A tutorial on support vector machines for pattern recognition. Data Min. Knowl. Disc., 2: Celebi, M.E., H.E. Kingravi, B. Uddin, H. Iyatomi, Y. Aslandogan, W.V. Stoecker, R. Moss, A methodological approach to the classification of dermoscopy images. Comput. Med. Imag. Graphics, 31(6): Celebi, M.E., H. Iyatomi, W. Stoecker, R.H. Moss, H. Rabinovitz, H.P. Soyer, Automatic detection of blue-white veil and related structures in dermoscopy images. Comput. Med. Imag. Graph., 32(8):